What Survivorship Bias Does to a Backtest

Survivorship bias is what happens when your historical universe contains only the companies that still exist today. Pull "all S&P 500 stocks" from a free API and backtest a strategy over 2005–2025, and you have silently conditioned on twenty years of future information: every company in your sample was, by construction, healthy enough to still be listed in 2025. Lehman Brothers, Washington Mutual, Enron, Blockbuster, Circuit City, and the thousands of less famous bankruptcies, distressed delistings, and going-private transactions simply never appear. Your strategy gets graded on a universe where catastrophic failure was impossible.

The bias always flatters you, and it flatters some strategies far more than others:

• Value and mean-reversion strategies are hit hardest. "Buy stocks that fell 60%" looks brilliant when every stock in your dataset eventually recovered — because the ones that went to zero were deleted before you started.
• Small-cap and low-price screens concentrate exactly where delisting rates are highest, so the missing failures are over-represented in the very names your screen selects.
• Momentum strategies suffer less on the long side but get a free pass on shorts: the best short candidates in history are the ones that no longer exist.

The academic literature has measured the damage. The classic result is from Tyler Shumway and Vincent Warther, who showed that CRSP — the most carefully curated equity database in existence — was itself missing many delisting returns, and that the missing ones were not random: delistings for performance reasons (bankruptcy, failure to meet listing requirements) were missing far more often than delistings for neutral reasons like mergers. For Nasdaq stocks, their 1999 study estimated the average missing performance-related delisting return at roughly −55%, and found the resulting bias was about 4.7× larger than the bias previously documented for NYSE/AMEX stocks (where Shumway's earlier work suggested a −30% correction). A material part of the historical "size effect" — small stocks appearing to outperform — turned out to be an artifact of how delisted losers exit the data. Later work showed the same delisting-return problem distorts accounting-based anomalies too; Alpha Architect has a good practitioner summary in "Dealing with Delistings".

Read that carefully, because it implies something stronger than the usual warning: even including delisted stocks is not enough if the final, usually terrible, return of each delisted stock is recorded wrong or not at all. A dataset can contain Lehman's price history and still overstate your strategy's performance if your backtester silently drops the position at the last quoted price instead of realizing the loss.

Survivorship bias is one of a family of data-induced backtest failures — look-ahead bias and data snooping are its siblings — that we cover in Why Most Backtests Fail. The difference is that survivorship bias is purely a purchasing decision. You cannot fix it with methodology. You fix it by buying the right data, which is what the rest of this article is about.

What "Bias-Free" Actually Requires (It's More Than Delistings)

Vendors love the phrase "survivorship-bias-free," but it covers a spectrum. A genuinely point-in-time equity universe needs four things, and most products deliver only the first:

1. Delisted securities. Every stock that traded during your backtest window must be present, with its full price history up to its final trading day. This is the table-stakes definition every vendor means when they use the phrase.
2. Delisting returns (or final-value handling). What did a holder actually receive when the stock left the exchange? Merger consideration, exchange to OTC at a collapsed price, or zero in bankruptcy? CRSP records this explicitly as a delisting return. Most commercial vendors do not — so your backtest engine needs an explicit rule (e.g., liquidate at last traded price, or apply a haircut for performance delistings per Shumway).
3. Point-in-time index constituents. If your strategy trades "the S&P 500" or "the Russell 3000," you need the membership list as it stood on each historical date. Testing today's constituents over the past 20 years is survivorship bias re-imported through the back door — index committees systematically remove failing companies and add successful ones. Only a few vendors (Norgate, Sharadar for the S&P 500, CRSP/Compustat at the institutional level) sell this.
4. Point-in-time fundamentals and reference data. Companies restate earnings, change tickers, and re-use other companies' old tickers. If you rank stocks on fundamentals, you need the numbers as originally reported, not the restated versions: S&P Global's own point-in-time vs. lagged fundamentals study shows the same factor backtest looks significantly better on restated (non-PIT) data than on what investors could actually have known — their canonical example is Enron, whose original fraudulent 1997–2000 filings survive only in point-in-time databases.

Keep this four-item checklist in mind as you read the vendor sections. "Includes delisted stocks" is item 1 only.

The Vendors, One by One

CRSP — the academic gold standard

The Center for Research in Security Prices at the University of Chicago is the dataset virtually every published asset-pricing result is built on. NYSE coverage starts December 31, 1925, and Nasdaq coverage December 14, 1972. Every security carries a permanent identifier (PERMNO) that survives ticker changes, and — uniquely among the sources in this article — CRSP maintains an explicit delisting file with delisting codes and delisting returns, so item 2 on the checklist is handled natively. Paired with Compustat via the CRSP/Compustat link, it is the institutional reference for both prices and fundamentals.

Access and price: almost everyone gets CRSP through WRDS (Wharton Research Data Services), which serves 400+ subscribing institutions. If you are at a university, you very likely already have it for free through your library. Direct commercial licensing exists but is quote-based and priced for institutions, not individuals.

Caveats: daily data only (no intraday); updates are periodic rather than nightly, so it is a research dataset, not a production trading feed. And per Shumway & Warther above, even CRSP has historically missing delisting returns for some performance delistings — the standard academic fix is to impute −30% (NYSE/AMEX) or −55% (Nasdaq) where the delisting return is missing and the delisting code indicates poor performance.

Compustat / S&P Capital IQ Point-in-Time — fundamentals as they were reported

Prices are only half the problem; if your signal uses fundamentals, you need them point-in-time. S&P Global's Compustat Point-in-Time / snapshot products preserve every vintage of every filing — preliminary, as-first-reported, and each subsequent restatement — so a backtest on date t can use exactly what was public on date t. The S&P quantamental research note linked above quantifies how much performance evaporates when you switch a factor backtest from restated to point-in-time data. Price: quote-based, institutional; academics typically reach it through WRDS alongside CRSP.

Norgate Data — the retail systematic trader's default

Norgate Data is an Australian vendor that has become the default answer on trading forums whenever someone asks for affordable survivorship-bias-free US daily data, and the published pricing explains why. Their US stocks packages (12-month terms, USD):

The package structure makes the point of this whole article for us: the survivorship-bias-free tiers cost roughly twice the survivor-only tiers, because collecting and maintaining dead companies is the expensive part. Platinum and Diamond include delisted securities, OTC securities that were formerly exchange-listed, and — critically, checklist item 3 — historical index constituency for major US indices, so you can reconstruct "the Russell 3000 as of June 2009" instead of today's list. Data is delivered through their updater application with integrations for AmiBroker and other retail backtesting platforms plus a Python package, which is why it pairs naturally with the frameworks we compared in our backtesting-framework shootout.

Caveats: end-of-day only; US and Australian markets only; no delisting returns (your backtester decides what a final exit is worth); fundamentals coverage is thin compared with Sharadar — this is a price-series product.

Sharadar (Nasdaq Data Link) — prices plus fundamentals, API-first

Sharadar Equity Prices (SEP), distributed through Nasdaq Data Link (the former Quandl), covers adjusted and unadjusted end-of-day prices and corporate actions for 20,000+ US public companies back to 1998, active and delisted, with a companion fund-price database covering 6,000+ active and delisted ETFs and CEFs. The ticker metadata table flags delisted status explicitly, and the companion Core US Fundamentals database provides about 25 years of fundamentals that Sharadar describes as "nearly completely free from survivorship bias." A separate table tracks historical S&P 500 constituents — partial credit on checklist item 3 (the S&P 500, but not the Russell indices).

Price: subscriptions are per-database and monthly, but Nasdaq Data Link only displays the numbers after you create a (free) account and select a professional or non-professional license — so treat it as published-on-login rather than public. The same data is also resold inside QuantRocket if you want it pre-integrated with a backtester.

Caveats: history starts in 1998 (Norgate Diamond and CRSP go much deeper); end-of-day only; no delisting returns. Its real differentiator is being the cheapest credible way to get both a delisted-inclusive price history and point-in-time-ish fundamentals through one API.

QuantConnect / LEAN — free, if you live in their ecosystem

QuantConnect's hosted US Equities dataset (sourced from AlgoSeek) is explicitly survivorship-bias-free: every stock that traded on the US consolidated tape since January 1998, roughly 27,500 securities, at resolutions from tick to daily. The companion US Equity Security Master handles the reference-data half of the problem — splits, dividends, delistings, mergers, and ticker changes — automatically inside the LEAN engine, so a backtest position in a stock that delists is wound down by the engine rather than silently frozen.

Price: free for backtesting on their cloud platform. The catch is the boundary: the data lives inside QuantConnect. Using it locally with open-source LEAN means buying data through their market, and exporting it to your own pandas research stack is restricted by licensing. It is the best zero-budget option and the only one here where delisting handling (checklist item 2) is built into the execution engine rather than left to you.

FirstRate Data — survivorship-bias-free intraday bars

Everything above is daily data. If you need minute bars without survivorship bias, FirstRate Data's "Stocks Complete" bundle is one of very few affordable options: 1-minute through daily bars from January 2000 for 16,000+ tickers, of which over 7,000 are delisted, with both adjusted and unadjusted prices and pre/post-market trades included. It is sold as a one-time downloadable purchase (price on the bundle page) that includes one month of updates, with ongoing updates at $59.95/month thereafter.

Caveats: no index-constituent history, no fundamentals, no delisting returns — it is a bar-data archive, not a reference-data product. Pair it with Norgate or Sharadar metadata if you need to know why a ticker stopped trading.

EODHD — budget global coverage, with an asterisk on delistings

EODHD is a popular budget API (30+ years of US end-of-day history, 60+ exchanges globally) that does provide delisted-company data — but with a structural limitation their own documentation states plainly: for companies delisted after 2018, you get fundamentals, dividends/splits, EOD prices, and (from 2021) intraday data; for companies delisted before 2018, only end-of-day price history is available, and per-ticker availability should be confirmed with support. In other words, the delisted universe gets progressively thinner the further back you go — exactly where survivorship bias does the most damage.

Price: per the published pricing page, plans run from free (20 API calls/day, no delisted data) through EOD All World at $19.99/month, EOD+Intraday at $29.99/month, Fundamentals at $59.99/month, to the All-In-One package at $99.99/month. The plan pages don't clearly state which tier unlocks delisted coverage, so confirm before subscribing. Reasonable for global breadth on a budget; run the validation checks on the pre-2018 graveyard before trusting it for deep backtests.

Polygon.io (now Massive) — great API, verify the graveyard

Polygon.io, which rebranded as Massive in late 2025 (same APIs and products), is many developers' favorite US market data API, with stocks plans that have recently run in the $29–$199/month range across third-party roundups (the current tiers are on their pricing page). Its historical aggregates do include tickers that have since delisted, and because the data is organized per-ticker as traded, it doesn't retroactively delete failures.

The asterisk is reference data. Community projects that have tried to build full survivorship-bias-free databases from it — for example this open-source Polygon stock-database project — and reviews such as InvestingRobots' report that metadata for delisted tickers can be sparse: ticker records with no company name, missing trading-date ranges, or no classification. Price bars for dead companies are mostly there; knowing what those companies were, and screening them point-in-time, takes work. If Polygon/Massive is your platform, treat survivorship-bias-freedom as something you must assemble and verify, not something you bought.

Databento — as-traded microstructure data, bias-free by construction

Databento sells raw, as-traded market data straight from exchange feeds with usage-based pricing (plus flat-rate plans and a $125 trial credit). Because it captures the tape as it happened, every symbol that printed a trade is in the data — its security master spans 310,000+ listed and delisted instruments. There is no survivorship filtering to undo. The trade-offs run the other way: its consolidated US Equities Mini daily dataset only begins in March 2023 (venue-specific feeds like Nasdaq's go back further), so it is the right tool for microstructure and execution research — the territory of our LOBSTER and HFT dataset guide — not for a 25-year daily-universe backtest.

Bloomberg and LSEG/Refinitiv — quote-based, point-in-time at a price

For completeness: both Bloomberg (via Terminal/Data License) and LSEG (Refinitiv) offer delisted coverage and point-in-time products — LSEG has written publicly about using point-in-time data to avoid backtest bias. Pricing is quote-based and enterprise-scale in both cases. If your firm already pays for one of them, exhaust that entitlement before buying anything above; if you're an individual, these are not aimed at you.

Side-by-Side Comparison

How to Validate a Dataset Yourself

Never take "survivorship-bias-free" on faith — it takes ten minutes to check. Three sanity tests catch most problems:

Three notes on interpreting the results. First, symbology differs by vendor: bankrupt companies often traded their final months under a Q-suffixed OTC ticker (ENRNQ, WAMUQ, SHLDQ), some vendors keep the original symbol, Norgate namespaces delisted tickers separately, and CRSP sidesteps tickers entirely with PERMNOs — read the symbology docs before declaring a corpse missing. Second, check the universe size through time: the number of US listed companies peaked in the late 1990s and has roughly halved since, so a clean dataset shows a larger tradable universe in 1999 than today. If your 1999 universe is smaller than your 2024 universe, you are looking at survivors. Third, check final prices: a delisted-for-cause stock whose last recorded price is its pre-collapse level (rather than pennies) means your backtester will exit failures at fantasy prices — which is the delisting-return problem from section 1 wearing a different hat.

Recommendations by Budget

$0 — student, hobbyist, or just exploring: backtest inside QuantConnect. The hosted AlgoSeek data is survivorship-bias-free from 1998 and the engine handles delistings for you. The constraint is the walled garden, not the data quality. (Yahoo Finance and most free APIs are survivor-only; treat any free downloaded universe as biased until proven otherwise.)

Under $1,000/year — independent quant trading daily bars: Norgate Platinum ($630/yr) if your strategies are price-based and index-scoped; the point-in-time Russell/S&P constituent lists alone justify the price. Choose Sharadar instead if your signals need fundamentals through an API, and add Norgate later if you need pre-1998 history or non-S&P-500 constituent lists. EODHD is the budget pick for global coverage if you've validated that its pre-2018 delisted coverage is good enough for your window.

Intraday backtests: FirstRate Data's Stocks Complete bundle (7,000+ delisted tickers at 1-minute resolution from 2000) or QuantConnect minute data. For tick-level and order book research, Databento — but plan your universe-history strategy separately, since its consolidated daily history is short.

Academic: CRSP via WRDS, full stop. It's the only option with true delisting returns, your institution probably already pays for it, and referees will expect it. Apply the Shumway corrections for missing performance-delisting returns.

Institutional: CRSP + Compustat Point-in-Time for research; Bloomberg or LSEG entitlements you likely already have for production reference data; Databento or vendor tick archives for execution research. At this tier every price is quote-based — the negotiating leverage is knowing exactly which of the four checklist items each contract actually delivers.

Bottom Line

Survivorship bias is the one backtest killer you can solve with a purchase order, and the market has stratified cleanly: CRSP for academics, Norgate and Sharadar for serious retail quants at well under $1,000 a year, QuantConnect for free inside its sandbox, FirstRate for intraday, and quote-based institutional products above that. But "survivorship-bias-free" on a sales page only reliably means delisted price series included — delisting returns, point-in-time index membership, and as-reported fundamentals each cost extra or come from a different vendor entirely. Buy against the four-item checklist, then spend ten minutes counting corpses in the data before you trust a single equity curve built on it.

Related reading: Why Most Backtests Fail: Overfitting, Look-Ahead Bias, and Data Snooping • Handling Survivorship Bias in Alternative Data Backtests • Walk-Forward Analysis: Designing a Rolling Out-of-Sample Test